Are there any well-behaved POSIX interval timers?

Question

Inspired by the last leap second, I've been exploring timing (specifically, interval timers) using POSIX calls.

POSIX provides several ways to set up timers, but they're all problematic:

• sleep and nanosleep—these are annoying to restart after they're interrupted by a signal, and they introduce clock skew. You can avoid some, but not all, of this skew with some extra work, but these functions use the realtime clock, so this isn't without pitfalls.
• setitimer or the more modern timer_settime—these are designed to be interval timers, but they're per-process, which is a problem if you need multiple active timers. They also can't be used synchronously, but that's less of a big deal.
• clock_gettime and clock_nanosleep seem like the right answer when used with CLOCK_MONOTONIC. clock_nanosleep supports absolute timeouts, so you can just sleep, increment the timeout, and repeat. It's easy to restart after an interruption that way, too. Unfortunately, these functions might as well be Linux-specific: there's no support for them on Mac OS X or FreeBSD.
• pthread_cond_timedwait is available on the Mac and can work with gettimeofday as a kludgy workaround, but on the Mac it can only work with the realtime clock, so it's subject to misbehavior when the system clock is set or a leap second happens.

Is there an API I'm missing? Is there a reasonably portable way to create well-behaved interval timers on UNIX-like systems, or does this sum up the state of things today?

By well-behaved and reasonably portable, I mean:

• Not prone to clock skew (minus, of course, the system clock's own skew)
• Resilient to the system clock being set or a leap second occurring
• Able to support multiple timers in the same process
• Available on at least Linux, Mac OS X, and FreeBSD

A note on leap seconds (in response to R..'s answer):

POSIX days are exactly 86,400 seconds long, but real-world days can rarely be longer or shorter. How the system resolves this discrepancy is implementation-defined, but it's common for the leap second to share the same UNIX timestamp as the previous second. See also: Leap Seconds and What To Do With Them.

The Linux kernel leap second bug was a result of failing to do housekeeping after setting the clock back a second: https://lkml.org/lkml/2012/7/1/203. Even without that bug, the clock would have jumped backwards one second.

Answer 1

kqueue and kevent can be utilized for this purpose. OSX 10.6 and FreeBSD 8.1 add support for EVFILT_USER, which we can use to wake up the event loop from another thread.

Note that if you use this to implement your own condition and timedwait, you do not need locks in order to avoid race conditions, contrary to this excellent answer, because you cannot "miss" an event on the queue.

Sources:

• FreeBSD man page
• OS X man page
• kqueue tutorial
• libevent source code

Example Code

Compile with clang -o test -std=c99 test.c

#include <sys/types.h>
#include <sys/event.h>
#include <sys/time.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>
#include <pthread.h>

// arbitrary number used for the identifier property
const int NOTIFY_IDENT = 1337;

static int kq;

static void diep(const char *s) {
   perror(s);
   exit(EXIT_FAILURE);
}

static void *run_thread(void *arg) {
    struct kevent kev;
    struct kevent out_kev;
    memset(&kev, 0, sizeof(kev));
    kev.ident = NOTIFY_IDENT;
    kev.filter = EVFILT_USER;
    kev.flags = EV_ADD | EV_CLEAR;

    struct timespec timeout;
    timeout.tv_sec = 3;
    timeout.tv_nsec = 0;

    fprintf(stderr, "thread sleep\n");

    if (kevent(kq, &kev, 1, &out_kev, 1, &timeout) == -1)
        diep("kevent: waiting");

    fprintf(stderr, "thread wakeup\n");

    return NULL;
}

int main(int argc, char **argv) {
    // create a new kernel event queue
    kq = kqueue();
    if (kq == -1)
        diep("kqueue()");


    fprintf(stderr, "spawn thread\n");
    pthread_t thread;
    if (pthread_create(&thread, NULL, run_thread, NULL))
        diep("pthread_create");

    if (argc > 1) {
        fprintf(stderr, "sleep for 1 second\n");
        sleep(1);
        fprintf(stderr, "wake up thread\n");

        struct kevent kev;
        struct timespec timeout = { 0, 0 };

        memset(&kev, 0, sizeof(kev));
        kev.ident = NOTIFY_IDENT;
        kev.filter = EVFILT_USER;
        kev.fflags = NOTE_TRIGGER;

        if (kevent(kq, &kev, 1, NULL, 0, &timeout) == -1)
            diep("kevent: triggering");
    } else {
        fprintf(stderr, "not waking up thread, pass --wakeup to wake up thread\n");
    }

    pthread_join(thread, NULL);
    close(kq);
    return EXIT_SUCCESS;
}

Output

$ time ./test
spawn thread
not waking up thread, pass --wakeup to wake up thread
thread sleep
thread wakeup

real    0m3.010s
user    0m0.001s
sys 0m0.002s

$ time ./test --wakeup
spawn thread
sleep for 1 second
thread sleep
wake up thread
thread wakeup

real    0m1.010s
user    0m0.002s
sys 0m0.002s

Answer 2

POSIX timers (timer_create) do not require signals; you can also arrange for the timer expiration to be delivered in a thread via the SIGEV_THREAD notification type. Unfortunately glibc's implementation actually creates a new thread for each expiration (which both has a lot of overhead and destroys any hope of realtime-quality robustness) despite the fact that the standard allows reuse of the same thread for each expiration.

Short of that, I would just recommend making your own thread that uses clock_nanosleep with TIMER_ABSTIME and CLOCK_MONOTONIC for an interval timer. Since you mentioned that some broken systems might lack these interfaces, you could simply have a drop-in implementation (based e.g. on pthread_cond_timedwait) on such systems, and figure it might be lower-quality due to lack of monotonic clock, but that this is just a fundamental limitation of using a low-quality implementation like MacOSX.

As for your concern about leap seconds, if ntpd or similar is making your realtime clock jump backwards when a leap second occurs, that's a serious bug in ntpd. POSIX time (seconds since the epoch) are in units of calendar seconds (exactly 1/86400 of a day) per the standard, not SI seconds, and thus the only place leap second logic belongs on a POSIX system (if anywhere) is in mktime/gmtime/localtime when they convert between time_t and broken-down time. I haven't been following the bugs that hit this time, but they seem to have resulted from system software doing a lot of stupid and wrong stuff, not from any fundamental issue.

Answer 3

You can look at the question here for clock_gettime emulation, which I've also supplied an answer for, but helped me as well. I've recently added a simple timer to a little repository I keep for Mac OS X timing that partially emulates POSIX calls. A simple test runs the timer at 2000Hz. The repo is called PosixMachTiming. Try it out.

PosixMachTiming is based on Mach. It seems some of the timing-related Mach API has disappeared from Apple's pages and has deprecated, but there are still bits of source code floating around. It looks like AbsoluteTime units and kernel abstractions found here are the new way of doing things. Anyways the PosixMachTiming repo still works for me.

Overview of PosixMachTiming

clock_gettime is emulated for CLOCK_REALTIME by a mach function calls that tap into the system realtime clock, dubbed CALENDAR_CLOCK.

The clock_gettime is emulated for CLOCK_MONOTONIC by using a global variable (extern mach_port_t clock_port). This clock is initialized when the computer turns on or maybe wakes up. I'm not sure. In any case, it's the global variable that the function mach_absolute_time() calls.

clock_nanosleep(CLOCK_MONOTONIC, TIMER_ABSTIME, ...) is emulated by using nanosleep on the difference between current time and the absolute monotonic time.

itimer_start() and itimer_step() are based on calling clock_nanosleep for a target absolute monotonic time. It increments the target time by the time-step at each iteration (not the current time) so that clock skew is not an issue.

Note that this does not satisfy your requirement to be able to support multiple timers in the same process.

Answer 4

There's also the new CLOCK_TAI which doesn't take the leap second corrections the real time clocks do.

Answer 5

We can make use of timer_create () or timerfd_create () . Their examples are present in man page .